Model Reduction through Machine Learning Tools Using Simulation Data with High Variance

Enhancing the computational speed through model reduction can facilitate the use of a complex system model for a design task. A part of the system model that is demanding in numerical calculation can be replaced by a surrogate model using machine learning tools such as support vector machine (SVM) and artificial neural network (ANN), which may be an effective way to find a highly nonlinear regression model with a multi-dimensional input.

However, obtaining a proper dataset for training the model is one of the biggest challenges to find such a model, especially when the input of the model has a high dimension. In this regard, running a system simulation with a high-fidelity model in the possible operational mode can provide a relevant and appropriate size of the data-set with high variance.

In this paper, a propulsion simulation of a marine vessel system based on the first principle models is used to generate a synthetic data-set for training ANN and SVM models for a cylinder model of a main engine. The trained models were tested with a static data-set, an open-loop dynamic simulation, and a closed-loop dynamic simulation, and compared to the original models.

The results show that the outputs from the surrogate models agree well with the 0D models except some load increasing situation in the closed-loop dynamic simulation. The system model using the surrogate models showed an order-higher simulation speed than the 0D model, and the gap will increase as the rated shaft speed of the diesel engine increases.